We introduce ZGCM-1, a fully open 7‑billion‑parameter dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency.\ \ The core premise is that compact models cannot passively memorize the open web, but by coupling deliberate internal reasoning with active external tool use they can overcome parametric capacity limits and handle a 256K context window.\ \ To support this paradigm we develop an end‑to‑end high‑efficiency training recipe:\
- Architecture & system co‑design: interleaved gated sliding‑window combined with full attention, stabilized by an FP8 Muon optimizer.\
- Progressive curriculum & MDP mid‑training: context scaling across 16K, 64K, and 256K, reformulating interaction traces into Markov Decision Processes.\ \ We also establish an AI‑native R&D workflow where swarms of agents autonomously manage cluster operations, data curation, and rapid diagnostic evaluation.\ \ Extensive evaluations show ZGCM-1‑7B is competitive with other 7B models on general benchmarks, and remains on par with frontier models orders of magnitude larger (e.g., Qwen3‑235B‑A22B, GLM‑5.1) on challenging mathematical reasoning and agentic search suites.\ \ Our pre‑training design yields roughly a 4.2× improvement in time‑to‑loss for 16K context training, highlighting substantial efficiency gains.\ \ Across the full development lifecycle we distill eight actionable empirical findings covering architectural scaling, SFT quality pruning, long‑context generalization, and agentic co‑training dynamics.\ \ To enable community research, we open‑source model weights from pre‑training, mid‑training, and post‑training stages, intermediate checkpoints, training code, per‑stage data and recipes, as well as W&B logs.\ \ Review